Ensemble prediction of mitochondrial toxicity using machine learning technology

Mitochondria are intracellular organelles found in most eukaryotic cells. Mitochondrial function includes the generation of cellular energy, maintenance of cellular homeostasis, and metabolic processes. Mitochondrial impairment has increasingly been recognized as a contributor to drug-induced toxicity. We have developed predictive models using machine learning methods for the prediction of the in vitro outcome of the MMP and the GluGal assays. These models were built using the open-source software Flame, which supports the combination of models in ensembles and the extension of the training data with further experimental data, continuously improving the predictive power of the models Despite the large amount of available data for training of models, most compounds are evaluated as negative, resulting in an imbalanced class distribution. This paper demonstrates the application of a combination of equally distributed low-level models in an ensemble to account for the imbalance in the training set resulting in a classifier providing high sensitivity and specificity of 92 and 87% respectively. The model generated can further be integrated with mechanistic in vitro data for improved screening for mitochondrial toxicity.This research was funded from the Innovative Medicines Initiative 2 Joint Undertaking under grant agreements eTRANSAFE (777365). This Joint Undertaking receives support from the European Union’s Horizon 2020 research and innovation program and EFPIA companies in kind contribution